机器学习基石笔记：Homework #2 decision stump相关习题

原文地址：http://www.jianshu.com/p/4bc01760ac20

问题描述

程序实现

17-18

# coding: utf-8

import numpy as np
import matplotlib.pyplot as plt

def sign(n):
    if(n>0):
        return 1
    else:
        return -1

def gen_data():
    data_X=np.random.uniform(-1,1,(20,1))# [-1,1)
    data_Y=np.zeros((20,1))
    idArray=np.random.permutation([i for i in range(20)])
    for i in range(20):
        if(i<20*0.2):
            data_Y[idArray[i]][0]=-sign(data_X[idArray[i]][0])
        else:
            data_Y[idArray[i]][0] = sign(data_X[idArray[i]][0])
    data=np.concatenate((data_X,data_Y),axis=1)
    return data

def decision_stump(dataArray):
    minErrors=20
    min_s_theta_list=[]
    num_data=dataArray.shape[0]
    data=dataArray.tolist()
    data.sort(key=lambda x:x[0])
    for s in [-1.0,1.0]:
        for i in range(num_data):
            if(i==num_data-1):
                theta=(data[i][0]+1.0)/2
            else:
                theta=(data[i][0]+data[i+1][0])/2
            errors=0
            for i in range(20):
                pred=s*sign(data[i][0]-theta)
                if(pred!=data[i][1]):
                    errors+=1
            if(minErrors>errors):
                minErrors=errors
                min_s_theta_list=[]
            elif(minErrors<errors):
                continue
            min_s_theta_list.append((s, theta))
    i=np.random.randint(low=0,high=len(min_s_theta_list))
    min_s,min_theta=min_s_theta_list[i]
    return minErrors,min_s,min_theta

def computeEinEout(minErrors,min_s,min_theta):
    Ein=minErrors/20
    Eout=0.5+0.3*min_s*(abs(min_theta)-1)
    return Ein,Eout

if __name__=="__main__":
    Ein_list=[]
    Eout_list=[]
    for i in range(5000):
        dataArray=gen_data()
        minErrors,min_s,min_theta=decision_stump(dataArray)
        Ein,Eout=computeEinEout(minErrors,min_s,min_theta)
        Ein_list.append(Ein)
        Eout_list.append(Eout)

    # show results
    # 17 & 18
    print("the average Ein: ",sum(Ein_list)/5000)
    print("the average Eout: ",sum(Eout_list)/5000)

    plt.figure(figsize=(16,6))
    plt.subplot(121)
    plt.hist(Ein_list)
    plt.xlabel("Ein")
    plt.ylabel("frequency")
    plt.subplot(122)
    plt.hist(Eout_list)
    plt.xlabel("Eout")
    plt.ylabel("frequency")
    plt.savefig("EinEout.png")

19-20

# coding: utf-8

import numpy as np

def read_data(dataFile):
    with open(dataFile, 'r') as file:
        data_list = []
        for line in file.readlines():
            line = line.strip().split()
            data_list.append([float(l) for l in line])
        data_array = np.array(data_list)
        return data_array

def predict(s,theta,dataX):
    num_data=dataX.shape[0]
    res=s*np.sign(dataX-theta)
    return res

def decision_stump(dataArray):
    min_s_theta_list=[]
    num_data=dataArray.shape[0]
    minErrors=num_data
    data=dataArray.tolist()
    data.sort(key=lambda x:x[0])
    dataArray=np.array(data)
    dataX=dataArray[:,0].reshape(num_data,1)
    dataY=dataArray[:,1].reshape(num_data,1)
    for s in [-1.0,1.0]:
        for i in range(num_data):
            if(i==num_data-1):
                theta=(dataX[i][0]*2+1)/2
            else:
                theta=(dataX[i][0]+dataX[i+1][0])/2
            pred=predict(s,theta,dataX)
            errors=np.sum(pred!=dataY)
            if(minErrors>errors):
                minErrors=errors
                min_s_theta_list=[]
            elif(minErrors<errors):
                continue
            min_s_theta_list.append((s, theta))
    i=np.random.randint(low=0,high=len(min_s_theta_list))
    min_s,min_theta=min_s_theta_list[i]
    return minErrors,min_s,min_theta

def best_of_best(candidate):
    candidate.sort(key=lambda x:x[1])
    counts=0
    for i in range(len(candidate)):
        if(candidate[i][1]!=candidate[0][1]):
            break
        counts+=1
    i=np.random.randint(low=0,high=counts)
    return candidate[i][0],candidate[i][1],candidate[i][2],candidate[i][3]

if __name__=="__main__":
    data_array=read_data("hw2_train.dat")
    num_data=data_array.shape[0]
    num_dim=data_array.shape[1]-1
    candidate=[]
    dataY=data_array[:,-1].reshape(num_data,1)
    for i in range(num_dim):
        dataX=data_array[:,i].reshape(num_data,1)
        min_errors,min_s,min_theta=decision_stump(np.concatenate((dataX,dataY),axis=1))
        candidate.append([i,min_errors,min_s,min_theta])
    min_id,min_errors,min_s,min_theta=best_of_best(candidate)
    print("the optimal decision stump:\n","s: ",min_s,"\ntheta: ",min_theta)
    print("the Ein of the optimal decision stump:\n",min_errors/num_data)

    test_array=read_data("hw2_test.dat")
    num_test=test_array.shape[0]
    testY=test_array[:,-1].reshape(num_test,1)
    num_dim=test_array.shape[1]-1
    testX=test_array[:,min_id].reshape(num_test,1)
    pred=predict(min_s,min_theta,testX)
    print("the Eout of the optimal decision stump by Etest:\n",np.sum(pred!=testY)/num_test)

运行结果

17-18

19-20